Estimation for hidden Markov random fields (Q1918169)

The idea of using a change of measure, under which the observation process becomes white noise, has been used extensively in continuous-time filtering theory where it gives rise to the Zakai equation. The technique is less well known in discrete time. The recent book of the authors and \textit{J. B. Moore} [Hidden Markov models. Estimation and control. (1995; Zbl 0819.60045)] exploits the method to estimate hidden Markov models. There it is indicated how the idea can be extended to hidden Markov random fields; this is developed in the present paper. A random field \(X\) indexed by a finite set \(L\) is considered. At each point \(\ell\) of the lattice \(X_\ell\) takes some value. The random field \(X\) is not directly observed; rather there is a noisy observation process \(Y\) which also takes one of the finite number of values at each \(\ell\in L\). Conditions are given which ensure \(X\) is a Markov random field. The problems discussed here are: given a set of observations \(\{Y_\ell, \ell \in L\}\), determine the most likely signal \(\{X_\ell,\;\ell\;\in L\}\) and, also, determine the parameters of the model, that is, the `transition probabilities' of the Markov random field \(X\) and the observations \(Y\). The paper is a presentation of a measure change parameter situation method for a random field; of course, the computational complexity required to implement these algorithms is significant.